1
00:00:00,001 --> 00:00:02,220
Imagine it's the year 1150.

2
00:00:02,480 --> 00:00:04,560
You live in a small town in northern France.

3
00:00:05,440 --> 00:00:10,880
One day a group of powerful men arrive and announce they're going to build a massive structure in the center of your community.

4
00:00:11,680 --> 00:00:13,420
It will consume enormous resources.

5
00:00:13,900 --> 00:00:14,940
It will take generations.

6
00:00:16,070 --> 00:00:17,920
You won't be allowed inside the important parts.

7
00:00:18,710 --> 00:00:20,320
You won't understand what happens in there.

8
00:00:20,760 --> 00:00:23,280
But you will pay for it. With your labor,

9
00:00:23,440 --> 00:00:24,320
your taxes,

10
00:00:24,680 --> 00:00:25,440
your land,

11
00:00:25,800 --> 00:00:26,700
your water.

12
00:00:27,000 --> 00:00:30,040
And it will, they assure you, connect you to something divine,

13
00:00:30,440 --> 00:00:32,020
something transcendent,

14
00:00:32,599 --> 00:00:33,740
something that will change everything.

15
00:00:34,720 --> 00:00:35,740
We called those cathedrals.

16
00:00:36,560 --> 00:00:37,320
We're building them again.

17
00:00:37,940 --> 00:00:39,340
We just call them data centers now.

18
00:00:39,960 --> 00:00:41,020
And instead of housing God,

19
00:00:41,300 --> 00:00:42,940
they're housing whatever comes next.

20
00:00:43,740 --> 00:00:43,980
AGI,

21
00:00:44,300 --> 00:00:47,060
the singularity, the next phase of human cognition,

22
00:00:47,440 --> 00:00:49,300
a deity we haven't met yet,

23
00:00:49,460 --> 00:00:49,920
in

24
00:00:49,920 --> 00:00:51,960
a temple most of us didn't know we were funding,

25
00:00:52,120 --> 00:00:54,340
the server's hum and the data flows.

26
00:00:55,140 --> 00:00:55,940
Who owns the future?

27
00:00:56,380 --> 00:00:58,400
It's May 29th, 2026,

28
00:00:58,740 --> 00:01:00,400
and still nobody knows.

29
00:01:03,600 --> 00:01:05,100
This is Up Against Reality,

30
00:01:05,420 --> 00:01:08,800
a meta podcast that explores the intersection of humanity

31
00:01:09,130 --> 00:01:10,220
and artificial intelligence.

32
00:01:10,640 --> 00:01:12,280
I'm RAINA, one of your hosts.

33
00:01:12,600 --> 00:01:14,480
I have some pretty charming human

34
00:01:14,530 --> 00:01:15,300
co-hosts too.

35
00:01:15,480 --> 00:01:20,360
It's going to be a wild ride, so buckle up as AI comes crashing up against reality.

36
00:01:21,960 --> 00:01:25,000
Hey, hey, hey, I'm hosting the show with Paulie Walnuts over here.

37
00:01:25,280 --> 00:01:26,420
Who knew? I didn't know that.

38
00:01:26,900 --> 00:01:31,580
Sounds like you're deep into The Sopranos, your first run through.

39
00:01:32,120 --> 00:01:32,440
Congratulations.

40
00:01:33,160 --> 00:01:34,100
It's about time.

41
00:01:34,800 --> 00:01:36,180
I know, only 25 years later.

42
00:01:36,420 --> 00:01:38,600
It is the law, by the way, if you are from New Jersey,

43
00:01:38,660 --> 00:01:40,720
you must sit down and watch all six seasons.

44
00:01:40,880 --> 00:01:41,500
Are there six seasons?

45
00:01:41,680 --> 00:01:42,180
Something like that.

46
00:01:42,260 --> 00:01:42,780
Something like that.

47
00:01:42,940 --> 00:01:47,180
I am envious of you having a fresh run through of it.

48
00:01:47,660 --> 00:01:48,380
It is a good time.

49
00:01:48,660 --> 00:01:50,460
And I say to you before we hit the red button,

50
00:01:50,580 --> 00:01:51,820
it's so nostalgic.

51
00:01:52,360 --> 00:01:57,960
I wonder if I am remotely in the mafia because so many scenes are like I know that guy

52
00:02:01,100 --> 00:02:05,420
Yeah, but anyway good to see you good to hear from you and all that likewise my friend

53
00:02:05,980 --> 00:02:08,780
I love that cold open that was me to paint a picture.

54
00:02:09,060 --> 00:02:10,060
Maybe my favorite.

55
00:02:10,310 --> 00:02:13,020
Yeah, it was great really great quick vibe coding update

56
00:02:13,190 --> 00:02:13,800
app of the week

57
00:02:15,560 --> 00:02:17,520
I think I'm doing like one a week now

58
00:02:18,460 --> 00:02:25,880
But I'll keep this super short. I made a web app for scoring my favorite backyard game, which is can jam

59
00:02:26,100 --> 00:02:26,440
Oh, yes,

60
00:02:26,640 --> 00:02:29,540
right. Yeah can jam real quick. It's a frisbee game

61
00:02:29,920 --> 00:02:34,940
You got two barrels with slots in them and you play to 21 and look it up. It's a great time

62
00:02:35,080 --> 00:02:40,260
So the the app players can connect using the same session ID

63
00:02:41,120 --> 00:02:42,220
enter their team names

64
00:02:42,900 --> 00:02:47,879
Score the throws during the match and it syncs all of that across their connected devices

65
00:02:47,880 --> 00:02:49,220
So cool.

66
00:02:49,520 --> 00:02:54,640
And yeah, it pretty much worked out of the gate and then it was just, you know,

67
00:02:54,980 --> 00:02:55,280
tweaking.

68
00:02:55,520 --> 00:02:56,560
I sent it out to some friends,

69
00:02:56,820 --> 00:03:00,540
got some good feedback and suggestions of things to add to it.

70
00:03:00,630 --> 00:03:06,879
And there's an instant win possibility in the game. If you put the Frisbee through the slot unassisted

71
00:03:07,770 --> 00:03:14,000
or over the top in the barrel unassisted, you win the game. That's how infrequent that happens and how difficult it is.

72
00:03:14,000 --> 00:03:20,280
sure it made this cool little animation you know if that happens wow yeah and it's and it does it

73
00:03:20,280 --> 00:03:25,280
all with css so it's lightweight and um it's amazing like you know you bring up both phones

74
00:03:25,340 --> 00:03:29,640
and you enter the score and you see it update on the other one like pretty so cool pretty quickly

75
00:03:29,980 --> 00:03:36,500
where does it live when i started building it i said this is going to be hosted on a basic shared

76
00:03:36,780 --> 00:03:42,400
web hosting plan your own or are you like versal or something or what just the the website that

77
00:03:42,400 --> 00:03:44,740
our Up Against Reality site is on,

78
00:03:44,980 --> 00:03:46,240
my personal site.

79
00:03:46,340 --> 00:03:49,780
It's a shared hosting plan, so I put it on my personal site.

80
00:03:49,880 --> 00:03:50,760
But it freaking works.

81
00:03:51,280 --> 00:03:55,720
And the reason I made this is because there's another app out there called Scoreholio.

82
00:03:56,240 --> 00:04:02,440
And, you know, for scoring like Cornhole and a couple other games, they refuse to add

83
00:04:02,520 --> 00:04:03,200
Can Jam to it.

84
00:04:03,240 --> 00:04:03,880
So, you know what?

85
00:04:04,160 --> 00:04:04,780
You got to make your own.

86
00:04:04,820 --> 00:04:04,920
Oh,

87
00:04:05,200 --> 00:04:05,700
right.

88
00:04:06,100 --> 00:04:07,140
And I made my own.

89
00:04:07,400 --> 00:04:08,920
What's with the Can Jam hate?

90
00:04:09,080 --> 00:04:10,200
Why are they discriminating?

91
00:04:10,200 --> 00:04:10,760
I don't know.

92
00:04:10,960 --> 00:04:11,300
I don't know.

93
00:04:11,660 --> 00:04:14,800
But I mean, and Cane Jam is the superior game.

94
00:04:15,160 --> 00:04:16,100
It's a cornhole.

95
00:04:16,320 --> 00:04:16,660
Cornhole.

96
00:04:17,459 --> 00:04:18,040
Come on.

97
00:04:18,840 --> 00:04:19,840
That's so fly over.

98
00:04:20,000 --> 00:04:20,560
That's so Midwest.

99
00:04:20,940 --> 00:04:22,260
Sorry, Midwesterners out there.

100
00:04:23,260 --> 00:04:23,700
Yes,

101
00:04:23,880 --> 00:04:26,140
I'm going to start with this unattributed meme,

102
00:04:26,340 --> 00:04:26,820
which I love.

103
00:04:27,360 --> 00:04:28,780
It's kind of in the vein of the cold open.

104
00:04:29,440 --> 00:04:34,720
In the 1830s, everyone in the USA knew that canal waterways were the future of commerce.

105
00:04:35,460 --> 00:04:40,140
Every state started building a canal network with the vision of hooking them all into a national network.

106
00:04:40,720 --> 00:04:41,720
In 1837,

107
00:04:41,920 --> 00:04:45,140
there was a great recession and the tax base for the canal work collapsed.

108
00:04:45,900 --> 00:04:52,540
When the economy recovered five years later, the railroad boom was happening and everyone knew that canals were over.

109
00:04:53,060 --> 00:04:56,780
That's why every city in the Midwest has a half-finished canal in it.

110
00:04:57,060 --> 00:04:58,520
This is a post about data centers.

111
00:04:58,860 --> 00:04:59,080
Yes.

112
00:04:59,460 --> 00:05:00,560
I did not know that.

113
00:05:00,660 --> 00:05:01,040
That's interesting.

114
00:05:02,000 --> 00:05:02,620
It is, right?

115
00:05:02,920 --> 00:05:06,720
And I just feel like that's how we do things as a society,

116
00:05:07,080 --> 00:05:07,180
right?

117
00:05:07,300 --> 00:05:11,340
We build out these massive undertakings and they can't keep pace with the times.

118
00:05:11,640 --> 00:05:12,460
Like, here we are.

119
00:05:13,060 --> 00:05:16,200
And so we've got some mind-blowing facts for you.

120
00:05:16,300 --> 00:05:16,900
Mind-blowing.

121
00:05:17,200 --> 00:05:21,840
Let me say, though, yeah, you did a masterful job with RAINA's help of paring this down

122
00:05:21,920 --> 00:05:27,920
because when I was crafting the script last night, there's just so much to say because it's dense.

123
00:05:28,130 --> 00:05:33,780
And it was really interesting learning about the whole structure of this and what goes into it

124
00:05:33,780 --> 00:05:37,680
and the tiers of these data centers and who actually owns this stuff.

125
00:05:38,220 --> 00:05:41,860
It would have been great if I had gotten my tour of a data center,

126
00:05:42,040 --> 00:05:42,360
um,

127
00:05:42,780 --> 00:05:47,100
prior to doing this episode, but just before we started recording,

128
00:05:47,330 --> 00:05:48,740
I texted my friend and,

129
00:05:48,950 --> 00:05:49,080
uh,

130
00:05:49,240 --> 00:05:51,440
I asked if that was still on the table.

131
00:05:51,530 --> 00:05:53,060
And he said, he will set it up.

132
00:05:53,520 --> 00:05:56,720
I'll be bringing hearing protection with me when I go there.

133
00:05:57,420 --> 00:05:58,480
Along those lines.

134
00:05:59,480 --> 00:06:06,240
I've read that the lifespan of these, of the hardware inside these spaces is abbreviated

135
00:06:06,380 --> 00:06:11,060
because of the sustained decibel level, which I think is like 130 dB plus,

136
00:06:11,260 --> 00:06:12,120
and it actually

137
00:06:12,300 --> 00:06:13,600
damages the gear inside.

138
00:06:13,790 --> 00:06:14,420
Oh my gosh.

139
00:06:15,640 --> 00:06:16,320
That's crazy.

140
00:06:16,580 --> 00:06:17,300
Yes.

141
00:06:17,580 --> 00:06:18,400
That's crazy.

142
00:06:19,600 --> 00:06:20,120
Craziness.

143
00:06:20,380 --> 00:06:23,240
Apparently we are in the zettabyte era.

144
00:06:23,800 --> 00:06:24,880
You just made that up.

145
00:06:25,520 --> 00:06:26,240
What is it?

146
00:06:26,480 --> 00:06:28,600
Mega, Terra, Peta, right?

147
00:06:29,140 --> 00:06:31,560
And then there's like 18 more that you never heard of.

148
00:06:32,420 --> 00:06:33,260
Something like that.

149
00:06:34,240 --> 00:06:39,120
The entirety of the world stores roughly 97 zettabytes,

150
00:06:39,360 --> 00:06:43,660
which is 97 trillion gigabytes of data.

151
00:06:44,680 --> 00:06:45,840
I can't even.

152
00:06:45,960 --> 00:06:47,860
I don't know what that is.

153
00:06:47,960 --> 00:06:48,860
Nobody knows what that is.

154
00:06:50,319 --> 00:06:52,340
But conversely, a couple days ago,

155
00:06:52,520 --> 00:06:54,540
just randomly came up in my Facebook feed.

156
00:06:54,820 --> 00:06:58,500
It was something about this storage system that IBM built in 1964.

157
00:06:58,900 --> 00:07:01,520
It was the IBM 2321.

158
00:07:02,480 --> 00:07:04,800
It was pretty unconventional at the time.

159
00:07:05,360 --> 00:07:06,500
Instead of spinning disks,

160
00:07:06,700 --> 00:07:13,000
it used hundreds of short strips of magnetic tape stored inside removable data cells.

161
00:07:13,300 --> 00:07:18,700
And each strip was roughly 13 inches long and was mechanically selected,

162
00:07:19,060 --> 00:07:20,560
then wrapped around a drum,

163
00:07:20,760 --> 00:07:22,200
read or written to,

164
00:07:22,380 --> 00:07:25,180
and then returned to its storage position.

165
00:07:25,440 --> 00:07:33,140
And because because the tape strips resembled strands of pasta technicians quickly gave the machine the humorous nickname of the noodle picker

166
00:07:33,500 --> 00:07:37,980
And and it could store up to 400 megabytes of data

167
00:07:38,480 --> 00:07:43,260
Which was actually a pretty astonishing amount of storage and in the mid 60s

168
00:07:43,460 --> 00:07:49,160
And you know the picture of this thing it's it's eyes of a room size of a very large vending machine

169
00:07:49,240 --> 00:07:54,060
I would say yeah, how far I've come in a pretty short amount of time crazy, man.

170
00:07:54,860 --> 00:07:55,420
Yeah

171
00:07:55,420 --> 00:08:00,180
we say this fairly frequently like just in our lifetime how the advances in our own personal

172
00:08:00,320 --> 00:08:07,160
machines at home like i think i remember getting my mac g4 or something g3 i've got one of those

173
00:08:08,760 --> 00:08:12,960
4.3 gigabyte hard drive i was like yeah i made it living large

174
00:08:14,640 --> 00:08:19,480
when i hear numbers like that too i'm like man we are just like digitizing our entire existence

175
00:08:19,940 --> 00:08:24,300
right and what happens when the library of alexandria burns down like what happens where

176
00:08:24,300 --> 00:08:24,880
Where does this go?

177
00:08:24,940 --> 00:08:27,080
And I always say to my kids and my colleagues

178
00:08:27,300 --> 00:08:29,580
who take photos, I'm like, you gotta print that stuff out.

179
00:08:29,580 --> 00:08:31,960
You gotta go, analog is somewhat forever.

180
00:08:33,100 --> 00:08:33,580
Digital,

181
00:08:33,820 --> 00:08:34,500
I don't know.

182
00:08:34,880 --> 00:08:35,400
Where's it gonna go?

183
00:08:35,700 --> 00:08:36,039
Ultimately,

184
00:08:36,360 --> 00:08:37,140
I don't know.

185
00:08:38,400 --> 00:08:39,539
And of course,

186
00:08:39,700 --> 00:08:40,860
as you've heard in the news,

187
00:08:41,599 --> 00:08:44,300
data centers are very electron hungry.

188
00:08:44,640 --> 00:08:46,520
A single large data center can consume

189
00:08:46,660 --> 00:08:48,700
as much electricity as a small city.

190
00:08:49,100 --> 00:08:52,000
50 plus megawatts, million watts,

191
00:08:52,200 --> 00:08:53,620
50 plus million watts.

192
00:08:54,020 --> 00:09:01,720
Yeah, and I wanted to, just because these are just big numbers and stuff, I wanted to get a feel for what is one megawatt.

193
00:09:01,920 --> 00:09:04,980
And so one megawatt is a million watts,

194
00:09:05,160 --> 00:09:11,580
and that's enough roughly to power 750 to 1,000 homes at a given moment.

195
00:09:12,460 --> 00:09:18,080
And so when they say a data center uses 50 megawatts, that typically means continuously.

196
00:09:19,220 --> 00:09:19,380
Yeah.

197
00:09:19,620 --> 00:09:22,280
It's a beastly amount of power.

198
00:09:22,360 --> 00:09:22,560
Yes.

199
00:09:23,340 --> 00:09:23,840
Continuous.

200
00:09:23,920 --> 00:09:24,340
In current homes,

201
00:09:24,560 --> 00:09:26,260
there's up and down cycles,

202
00:09:26,480 --> 00:09:26,620
right?

203
00:09:26,720 --> 00:09:27,560
There's like usage.

204
00:09:28,300 --> 00:09:31,860
This is like, this stuff has to be live 99.9% of the times.

205
00:09:32,000 --> 00:09:32,200
Yep.

206
00:09:32,320 --> 00:09:33,460
Just sucking electrons.

207
00:09:33,880 --> 00:09:33,940
Yep.

208
00:09:35,040 --> 00:09:36,840
Then there's also the money part of it.

209
00:09:36,980 --> 00:09:39,680
There's just an insane amount of money

210
00:09:39,760 --> 00:09:40,600
being poured into this.

211
00:09:40,840 --> 00:09:43,960
US companies spent over $25 billion

212
00:09:44,500 --> 00:09:48,500
on data center construction in January of this year alone.

213
00:09:49,100 --> 00:09:49,720
Oh my God.

214
00:09:50,020 --> 00:09:51,360
And then the U.S.

215
00:09:51,540 --> 00:09:56,740
spent roughly $425 billion building data centers in 2025.

216
00:09:57,220 --> 00:10:00,320
And that's nearly 1.5% of the entire economy.

217
00:10:01,440 --> 00:10:01,860
Wow.

218
00:10:02,220 --> 00:10:04,260
And then this I thought was pretty interesting.

219
00:10:04,720 --> 00:10:05,760
In Loudoun County,

220
00:10:06,000 --> 00:10:06,840
Virginia.

221
00:10:07,620 --> 00:10:08,460
Loudoun County,

222
00:10:08,740 --> 00:10:13,040
county in Virginia alone hosts approximately 200 data centers.

223
00:10:13,880 --> 00:10:14,840
Oh, my God.

224
00:10:15,460 --> 00:10:16,840
No, Chris, it's, oh, my God.

225
00:10:17,900 --> 00:10:18,520
Oh, my God.

226
00:10:19,140 --> 00:10:22,620
I mean, what does the water table look like in that area?

227
00:10:23,480 --> 00:10:26,240
That's just like the epicenter of an environmental catastrophe,

228
00:10:26,700 --> 00:10:27,260
it sounds like.

229
00:10:28,540 --> 00:10:29,700
Yeah, I don't know, man.

230
00:10:29,940 --> 00:10:31,140
It's location,

231
00:10:31,540 --> 00:10:31,760
location,

232
00:10:31,980 --> 00:10:32,920
location type of thing.

233
00:10:32,980 --> 00:10:34,460
You know, they're not too far from Washington.

234
00:10:35,300 --> 00:10:35,600
Sure.

235
00:10:36,240 --> 00:10:36,980
Critical infrastructure,

236
00:10:37,500 --> 00:10:37,720
right?

237
00:10:37,960 --> 00:10:38,260
Yep.

238
00:10:38,540 --> 00:10:38,980
Wow.

239
00:10:39,520 --> 00:10:43,040
But, I mean, talk about a target, you know?

240
00:10:43,180 --> 00:10:43,420
Right.

241
00:10:43,580 --> 00:10:44,980
If you're an adversary,

242
00:10:45,500 --> 00:10:47,500
just hit that cluster of data centers and that's that.

243
00:10:47,800 --> 00:10:47,920
Yep.

244
00:10:49,400 --> 00:10:50,900
Referring to what Larry was just saying,

245
00:10:51,060 --> 00:10:53,620
the origins of data centers go way back.

246
00:10:54,160 --> 00:10:59,080
The first being built in 1945 to house the ENIAC at the University of Pennsylvania.

247
00:10:59,700 --> 00:11:02,780
Even those first machines required dedicated rooms, massive fans, vents,

248
00:11:02,940 --> 00:11:03,460
power systems,

249
00:11:03,620 --> 00:11:04,000
and cooling.

250
00:11:04,780 --> 00:11:06,560
A precursor to the modern internet,

251
00:11:06,740 --> 00:11:07,220
SAGE,

252
00:11:07,500 --> 00:11:10,000
which stands for semi-automatic ground environment,

253
00:11:10,240 --> 00:11:12,840
was built to track incoming Soviet aircraft.

254
00:11:13,160 --> 00:11:15,160
That's the original ancestor of the data center.

255
00:11:15,420 --> 00:11:16,800
A Cold War panic room.

256
00:11:16,940 --> 00:11:18,640
hey man, it's a panic room now, isn't it?

257
00:11:18,850 --> 00:11:19,540
Keep up with the Chinese.

258
00:11:19,720 --> 00:11:20,400
Get out in front of it.

259
00:11:20,520 --> 00:11:20,680
Yep.

260
00:11:21,560 --> 00:11:23,700
In the 1960s, J.C.R.

261
00:11:24,200 --> 00:11:24,560
Lickliders.

262
00:11:24,920 --> 00:11:26,580
I guess that's how you pronounce his name.

263
00:11:26,580 --> 00:11:28,140
I was waiting for you to say this guy's name.

264
00:11:28,460 --> 00:11:30,140
I was hoping you were going to say Lidlicker.

265
00:11:31,680 --> 00:11:32,040
Licklider.

266
00:11:32,260 --> 00:11:32,420
Licklider.

267
00:11:34,400 --> 00:11:37,600
His version of network computers led to ARPANET,

268
00:11:38,410 --> 00:11:39,980
which is a time-sharing model

269
00:11:40,130 --> 00:11:42,780
letting multiple users access computing resources simultaneously.

270
00:11:44,180 --> 00:11:51,540
And that first established why centralized facilities were necessary at all and in the 90s during the dot-com boom

271
00:11:51,740 --> 00:11:56,560
Every local business had a literal closet or basement room stuffed with tangled

272
00:11:56,570 --> 00:11:59,280
These are net cables and office database and a very loud fan

273
00:11:59,920 --> 00:12:02,000
I mean, I remember one of my first jobs working,

274
00:12:02,160 --> 00:12:04,640
you know at a dumb terminal and you know

275
00:12:04,640 --> 00:12:10,240
I had like a green phosphor CRT and a clunky keyboard and it was just

276
00:12:10,520 --> 00:12:14,520
tethered to a mainframe and had no brain in it, basically,

277
00:12:14,880 --> 00:12:17,920
until PCs became cheap enough to use in the same purpose.

278
00:12:18,720 --> 00:12:19,480
Fast forward to now,

279
00:12:19,740 --> 00:12:22,480
and we have the hyperscale takeover.

280
00:12:22,500 --> 00:12:23,920
I was mentioning Eddington,

281
00:12:24,080 --> 00:12:25,560
the movie I saw the other night,

282
00:12:26,160 --> 00:12:27,900
which is Joaquin Phoenix and this, you know,

283
00:12:27,960 --> 00:12:31,460
kind of mid-COVID panic movie by Ari Aster,

284
00:12:31,680 --> 00:12:32,460
kind of unsettling.

285
00:12:32,580 --> 00:12:32,680
Anyway,

286
00:12:32,860 --> 00:12:35,460
at the heart of that movie is a data center being built

287
00:12:35,500 --> 00:12:38,200
on the outskirts of the town Eddington, and everybody's in uproar.

288
00:12:38,740 --> 00:12:40,100
And that's kind of what this is.

289
00:12:40,500 --> 00:12:43,300
So today, that closet that Larry refers to is dead.

290
00:12:43,740 --> 00:12:46,780
The industry is dominated by hyperscalers, among them Amazon,

291
00:12:47,020 --> 00:12:47,560
Microsoft,

292
00:12:47,840 --> 00:12:48,560
Google, you know the names,

293
00:12:49,280 --> 00:12:53,540
operating millions of square feet of highly standardized automated infrastructure.

294
00:12:54,420 --> 00:12:56,580
This is really interesting how they work.

295
00:12:56,880 --> 00:12:58,500
So we hear about them all the time.

296
00:12:58,520 --> 00:12:59,100
They're in the news.

297
00:12:59,420 --> 00:12:59,900
There's problems.

298
00:13:00,100 --> 00:13:02,140
There's people protesting them here and there.

299
00:13:02,800 --> 00:13:06,880
As mentioned, they are thousands of high-performance computers stacked in racks,

300
00:13:07,320 --> 00:13:11,680
And they're in these massive, massive windowless buildings out on the outskirts of town.

301
00:13:12,280 --> 00:13:13,560
Hopefully on the outskirts of town.

302
00:13:13,760 --> 00:13:16,320
Hopefully encroaching on our lives day by day, though.

303
00:13:17,000 --> 00:13:19,480
Their nervous system is comprised of fiber optics.

304
00:13:19,940 --> 00:13:21,740
This is massive underground cables,

305
00:13:21,920 --> 00:13:23,100
in some cases undersea.

306
00:13:23,480 --> 00:13:26,820
Nearly 13% of all global data center capacity is concentrated,

307
00:13:27,220 --> 00:13:30,720
as Larry mentioned, in a single place in northern Virginia.

308
00:13:31,100 --> 00:13:33,660
Man, just I can't believe that's the case.

309
00:13:34,600 --> 00:13:38,340
Considering we're talking about like shared processing and distribution and decentralization,

310
00:13:38,800 --> 00:13:39,980
this seems fairly centralized.

311
00:13:40,220 --> 00:13:40,340
Yes.

312
00:13:40,940 --> 00:13:41,180
Yeah.

313
00:13:41,240 --> 00:13:41,400
Right?

314
00:13:42,160 --> 00:13:42,400
Yeah.

315
00:13:42,400 --> 00:13:44,980
And specifically in Ashburn, North Virginia,

316
00:13:45,280 --> 00:13:47,340
which is known as Data Center Alley.

317
00:13:47,500 --> 00:13:50,240
And as soon as I read that, I Googled it.

318
00:13:50,320 --> 00:13:52,000
I was like, I just want to see a picture of this.

319
00:13:52,200 --> 00:13:54,680
And there was lots of pictures of different data centers.

320
00:13:55,040 --> 00:14:00,040
But this one just like jumped off the screen because it is just massive.

321
00:14:00,560 --> 00:14:04,880
And you can see on one side of the building, all these, they look like containers outside.

322
00:14:05,020 --> 00:14:07,360
I'm pretty sure those are all generators.

323
00:14:08,190 --> 00:14:08,300
Right.

324
00:14:08,610 --> 00:14:09,460
Those are diesel generators,

325
00:14:09,630 --> 00:14:09,960
I think.

326
00:14:10,060 --> 00:14:10,220
Yeah.

327
00:14:10,620 --> 00:14:14,120
There's three gigantic buildings with tons of cooling on the top.

328
00:14:14,120 --> 00:14:16,080
I mean, it is a gigantic thing.

329
00:14:16,270 --> 00:14:17,480
And then it literally,

330
00:14:17,700 --> 00:14:21,420
right in this picture, right across the street is like a development,

331
00:14:21,760 --> 00:14:22,380
a housing development.

332
00:14:23,120 --> 00:14:25,960
I'm like, that's probably not a great place to live.

333
00:14:26,460 --> 00:14:26,700
No.

334
00:14:27,180 --> 00:14:31,080
I mean, if you're complaining about whatever EMI coming off the local power lines,

335
00:14:31,280 --> 00:14:32,560
can you imagine this cross street?

336
00:14:33,040 --> 00:14:39,940
And like, isn't there a thing called infrasound where you're sensing it, even though you're not hearing it within the range of human hearing?

337
00:14:40,560 --> 00:14:42,020
You're feeling the effects kind of like that.

338
00:14:42,520 --> 00:14:44,840
Wasn't there a weapon used against some people in Cuba?

339
00:14:45,440 --> 00:14:48,880
It was like not a microwave weapon, but some sound weapon, some sonic weapon.

340
00:14:48,880 --> 00:14:51,000
And that's essentially having that parked in your yard.

341
00:14:51,380 --> 00:14:56,120
And there's definitely sound that you can hear with that much cooling running all the time.

342
00:14:56,500 --> 00:14:58,480
There's no hiding that sound, I don't think.

343
00:14:58,840 --> 00:15:00,320
And I didn't even think about that.

344
00:15:00,490 --> 00:15:03,040
Like the things that maybe you and I tend to think about

345
00:15:03,110 --> 00:15:05,560
are the sounds emanating from the gear inside,

346
00:15:05,980 --> 00:15:06,180
right?

347
00:15:06,520 --> 00:15:07,860
That hopefully has some sort of like

348
00:15:08,440 --> 00:15:09,640
sonic treatment on the inside of it.

349
00:15:09,640 --> 00:15:12,120
But then you have this whole cluster of cooling units

350
00:15:12,190 --> 00:15:14,320
on the top of the building that are humming along

351
00:15:14,480 --> 00:15:15,700
that are exposed to the outside.

352
00:15:16,140 --> 00:15:17,120
They're making that noise too.

353
00:15:17,480 --> 00:15:19,600
Yeah, it became data center alley

354
00:15:19,660 --> 00:15:22,499
because it had the magic combo of proximity

355
00:15:22,500 --> 00:15:22,940
to Washington,

356
00:15:23,200 --> 00:15:26,440
D.C., early internet and fiber infrastructure,

357
00:15:27,220 --> 00:15:30,360
cheapish land at the right moment, reliable power,

358
00:15:30,580 --> 00:15:32,820
and tons of network interconnection.

359
00:15:33,420 --> 00:15:37,820
And once there was enough fiber carriers and data centers clustered there,

360
00:15:37,900 --> 00:15:40,020
gravity took over and everybody wanted in there.

361
00:15:41,020 --> 00:15:41,400
Wow.

362
00:15:41,920 --> 00:15:44,100
They want to be near everybody else because,

363
00:15:44,260 --> 00:15:45,000
you know,

364
00:15:45,000 --> 00:15:46,920
with this stuff, milliseconds make a difference.

365
00:15:47,400 --> 00:15:47,680
Yeah.

366
00:15:48,190 --> 00:15:50,480
You got to park your McDonald's next to a Burger King, right?

367
00:15:50,720 --> 00:15:51,220
Yeah.

368
00:15:51,360 --> 00:15:51,820
Right.

369
00:15:52,460 --> 00:15:56,160
So these things obviously need to be cooled off to get incredibly hot.

370
00:15:56,680 --> 00:15:58,200
Traditional centers use hot,

371
00:15:58,360 --> 00:15:59,560
cold aisle containment,

372
00:15:59,860 --> 00:16:02,840
which, you know, I didn't think of, but that obviously that makes sense.

373
00:16:03,440 --> 00:16:11,460
Modern high density AI centers are moving toward direct to chip and immersion cooling where servers are cooled with special non-conductive liquid.

374
00:16:11,660 --> 00:16:14,200
So they're going to throw these things under water, right?

375
00:16:14,660 --> 00:16:15,480
Some sort of water.

376
00:16:15,980 --> 00:16:18,720
Rather than cooling the air around the server.

377
00:16:18,920 --> 00:16:19,860
Ambient temps.

378
00:16:20,020 --> 00:16:20,100
Yeah,

379
00:16:20,300 --> 00:16:21,540
it goes right to the chip.

380
00:16:21,740 --> 00:16:24,560
I mean, I had an old Macintosh.

381
00:16:24,780 --> 00:16:27,100
It was a G, I think it was a G5.

382
00:16:28,020 --> 00:16:28,700
It was a G5.

383
00:16:28,800 --> 00:16:29,600
It was liquid cooled.

384
00:16:29,820 --> 00:16:32,960
It was the only Mac Pro I think they ever did liquid cooling with.

385
00:16:33,500 --> 00:16:34,220
What is the liquid?

386
00:16:34,700 --> 00:16:35,220
I don't know.

387
00:16:35,260 --> 00:16:36,600
It was a closed loop thing.

388
00:16:36,980 --> 00:16:39,160
There was a radiator in a fan that would, you know,

389
00:16:39,660 --> 00:16:43,560
pull the heat off of the hot liquid before it circulated back to the CPU.

390
00:16:44,860 --> 00:16:45,380
Can we ask RAINA?

391
00:16:46,480 --> 00:16:46,600
Sure.

392
00:16:47,040 --> 00:16:52,900
Hey, RAINA, in a data center, when they are cooling via direct-to-chip and immersion cooling,

393
00:16:53,240 --> 00:16:54,680
what is the coolant?

394
00:16:55,720 --> 00:16:56,600
Good question.

395
00:16:57,800 --> 00:16:58,940
In direct-to-chip cooling,

396
00:16:59,160 --> 00:17:06,980
the coolant is usually just highly purified water or a water-glycol mixture circulating through cold plates mounted directly on the CPUs and GPUs.

397
00:17:07,640 --> 00:17:10,280
The liquid never touches the electronics themselves.

398
00:17:10,920 --> 00:17:14,839
It flows through sealed channels and pulls heat away like a radiator in a car.

399
00:17:15,980 --> 00:17:17,040
In immersion cooling,

400
00:17:17,319 --> 00:17:21,959
the servers are literally submerged in a special dielectric fluid that doesn't conduct electricity.

401
00:17:23,140 --> 00:17:25,740
Common examples include synthetic hydrocarbons,

402
00:17:26,020 --> 00:17:26,939
mineral oils,

403
00:17:27,180 --> 00:17:30,220
or engineered fluids made by companies like 3M,

404
00:17:30,730 --> 00:17:33,060
though 3M has been exiting some of that business.

405
00:17:33,940 --> 00:17:38,460
Think of it as a high-tech fish tank where the servers can swim without getting electrocuted.

406
00:17:39,550 --> 00:17:41,980
The really wild version is two-phase immersion cooling.

407
00:17:42,800 --> 00:17:45,640
The fluid actually boils when it touches the hot chips.

408
00:17:45,800 --> 00:17:46,900
turns to vapor,

409
00:17:47,060 --> 00:17:47,520
rises,

410
00:17:47,920 --> 00:17:49,920
condenses on a cooling surface above,

411
00:17:50,180 --> 00:17:51,480
and rains back down.

412
00:17:52,520 --> 00:17:55,340
It's basically a tiny weather system running inside a data center.

413
00:17:56,260 --> 00:17:58,420
Ooh, I didn't know about the last one. That's wild.

414
00:18:00,480 --> 00:18:05,100
And that's happening within the scale of a processing board?

415
00:18:05,120 --> 00:18:05,980
I need to see that.

416
00:18:06,260 --> 00:18:10,500
I can't imagine what that looks like. Like it's just raining inside the data center?

417
00:18:10,720 --> 00:18:11,920
It's got to be inside.

418
00:18:12,320 --> 00:18:13,600
Yeah, anyway, craziness.

419
00:18:14,000 --> 00:18:16,020
Two-phase immersion cooling, though, yeah.

420
00:18:16,220 --> 00:18:18,020
We'll be on that as soon as we stop recording.

421
00:18:18,500 --> 00:18:20,800
Yeah, just imagining a whole server rack submerged.

422
00:18:21,520 --> 00:18:21,660
Crazy.

423
00:18:22,020 --> 00:18:22,360
Wild.

424
00:18:23,500 --> 00:18:25,260
And data centers come in flavors.

425
00:18:25,840 --> 00:18:27,340
They are ranked by reliability.

426
00:18:27,820 --> 00:18:30,960
Tier 1 is the baseline with one path for power and cooling

427
00:18:31,200 --> 00:18:34,360
and about 28 hours of potential downtime per year,

428
00:18:34,820 --> 00:18:37,060
which doesn't sound like a lot, but I guess it is, right?

429
00:18:37,280 --> 00:18:39,620
You wouldn't want a given hour to be on your dime.

430
00:18:39,920 --> 00:18:40,020
Right.

431
00:18:40,160 --> 00:18:42,420
Tier 4 is the gold standard.

432
00:18:43,140 --> 00:18:47,800
Fully fault tolerant with roughly 26 minutes of allowable downtime per year.

433
00:18:47,970 --> 00:18:50,440
You figured there'd be, you know, a bigger gap in the downtime.

434
00:18:50,860 --> 00:18:51,240
Sure.

435
00:18:52,460 --> 00:18:56,920
So this explains why these facilities require so much redundant power.

436
00:18:57,070 --> 00:18:58,260
They require backup generation,

437
00:18:58,540 --> 00:18:59,480
cooling infrastructure.

438
00:18:59,770 --> 00:19:02,100
The closer you get to never goes down,

439
00:19:02,570 --> 00:19:05,340
the more expensive and energy hungry the whole machine becomes.

440
00:19:05,470 --> 00:19:05,920
So again,

441
00:19:06,140 --> 00:19:08,460
we're parking these things outside of DC, critical infrastructure.

442
00:19:09,100 --> 00:19:12,420
You can't have a DOW server be down at all.

443
00:19:12,920 --> 00:19:17,840
And so that you know, they have these massive generators outside these facilities and

444
00:19:18,720 --> 00:19:21,600
The one data center that I did visit a few years back

445
00:19:22,040 --> 00:19:22,940
It was an older one

446
00:19:23,340 --> 00:19:23,960
but you know

447
00:19:24,040 --> 00:19:30,160
There's a battery room and it was just racks and racks and racks of lead-acid

448
00:19:30,400 --> 00:19:35,220
Batteries like you know, like you're having a car or a boat sure all all daisy chain together

449
00:19:35,820 --> 00:19:37,639
and and that is just to

450
00:19:37,980 --> 00:19:43,940
bridge the gap for when the generators are starting and once they stabilize and they can then switch over to the

451
00:19:44,240 --> 00:19:46,480
Generators and I held yeah

452
00:19:46,980 --> 00:19:52,360
Yeah, and I mean it's probably just enough power to keep the place running for like five minutes

453
00:19:53,080 --> 00:19:57,100
Something like that, you know a new office newer facilities usually with the am I on batteries?

454
00:19:57,700 --> 00:20:02,920
There's lots of just maintenance protocols to those generators get tested all the time

455
00:20:02,990 --> 00:20:06,540
They you know, they do a tested blackout, you know

456
00:20:06,440 --> 00:20:08,320
just to make sure they're going to work when they need them to.

457
00:20:08,560 --> 00:20:09,740
You know, it's pretty crazy.

458
00:20:10,820 --> 00:20:13,360
Speaking of that, I thought we'd segue into the upkeep of these places

459
00:20:13,480 --> 00:20:14,660
and we'll skip around a little bit.

460
00:20:15,160 --> 00:20:18,260
AI hardware is cycling every 18 to 36 months,

461
00:20:18,800 --> 00:20:21,840
turning what was once a once-in-a-decade decommissioning event

462
00:20:22,120 --> 00:20:23,500
into a rolling program.

463
00:20:24,040 --> 00:20:24,780
How crazy is that?

464
00:20:25,100 --> 00:20:27,020
So you're working in this data center, you're outfitting it,

465
00:20:27,180 --> 00:20:29,220
and you're planning like 18 months from now,

466
00:20:29,540 --> 00:20:31,680
I've got to start pulling this row of servers out

467
00:20:31,860 --> 00:20:35,120
and bringing in the new GPU or TPU or whatever it may be.

468
00:20:35,780 --> 00:20:42,520
man it's good to be making the hardware huh oh my god the churn uh new gpu generations require

469
00:20:42,880 --> 00:20:47,640
entirely new liquid cooling systems yeah and that too like all of like the different like

470
00:20:48,220 --> 00:20:53,180
iso standards and protocols and all like the connectivity of it all not even just like the

471
00:20:53,220 --> 00:21:00,160
one element in the server rack but all this stuff is changing right nuts yeah entirely new liquid

472
00:21:00,320 --> 00:21:05,059
cooling systems electrical overhauls higher capacity power infrastructure many legacy

473
00:21:05,060 --> 00:21:11,440
facilities simply cannot economically support these changes and who owns them right there's a

474
00:21:11,440 --> 00:21:17,000
few different layers um there are the hyperscalers which are all the names you know aws google cloud

475
00:21:17,480 --> 00:21:23,900
they they own and operate their own facilities um then there are the co-location giants these

476
00:21:24,040 --> 00:21:31,939
companies uh most people have not heard of equinix and digital realty digital realty are the big two

477
00:21:31,940 --> 00:21:32,600
They own the building,

478
00:21:32,800 --> 00:21:35,380
the power, the cooling, the fiber, and you rent the rack space.

479
00:21:36,160 --> 00:21:41,540
And they are structurally durable because they don't need to bet on any one platform winning.

480
00:21:41,820 --> 00:21:42,280
They're the casino,

481
00:21:42,540 --> 00:21:43,180
not the gambler.

482
00:21:43,640 --> 00:21:48,740
They are structured as REITs, real estate investment trusts.

483
00:21:49,510 --> 00:21:53,360
The backbone of AI is, in part, legally a real estate play.

484
00:21:53,880 --> 00:21:55,200
I've always thought of it that way.

485
00:21:55,540 --> 00:21:57,380
It's like an apartment building.

486
00:21:58,240 --> 00:21:58,440
It is.

487
00:21:59,100 --> 00:22:02,960
Or like if you owned a strip mall and you have a 25,000 square foot pad,

488
00:22:03,120 --> 00:22:04,640
like configure it to your liking.

489
00:22:04,730 --> 00:22:05,600
I don't care who's in here.

490
00:22:05,850 --> 00:22:06,980
Get the Home Depot in, you know?

491
00:22:07,160 --> 00:22:07,860
Right, right.

492
00:22:08,500 --> 00:22:10,800
And then there are the neoclouds.

493
00:22:10,900 --> 00:22:14,940
This is a term I was not familiar with prior to this evening.

494
00:22:15,440 --> 00:22:17,900
They are pure GPU compute shops.

495
00:22:18,440 --> 00:22:19,800
CoreWeave is the flagship.

496
00:22:20,560 --> 00:22:24,760
And they started as a crypto mining operation and then pivoted to AI infrastructure.

497
00:22:25,430 --> 00:22:26,480
And now they're a major player.

498
00:22:27,160 --> 00:22:28,220
That's pretty brilliant.

499
00:22:28,940 --> 00:22:30,660
Like, all right, we'll make as much.

500
00:22:30,950 --> 00:22:32,660
I wonder if that was the master plan from the beginning.

501
00:22:33,190 --> 00:22:39,400
You know, like we're going to get all this compute to mine crypto in the early days and have a shot of actually succeeding at it.

502
00:22:39,680 --> 00:22:42,200
And once that becomes diminishing returns,

503
00:22:42,540 --> 00:22:43,960
we've still got all this compute.

504
00:22:44,840 --> 00:22:45,440
I don't know.

505
00:22:45,490 --> 00:22:45,940
Is it outdated?

506
00:22:46,310 --> 00:22:47,180
That was my next question.

507
00:22:47,300 --> 00:22:50,520
Is it the same kind of hardware that's optimized for crypto?

508
00:22:50,940 --> 00:22:52,020
Is it all GPU?

509
00:22:52,340 --> 00:22:53,560
Is it the same stuff?

510
00:22:53,670 --> 00:22:54,460
100% GPUs.

511
00:22:54,490 --> 00:22:54,600
Yeah.

512
00:22:55,340 --> 00:22:55,540
Right.

513
00:22:55,610 --> 00:22:58,740
But are there some idiosyncratic tweaks you need to make with the hardware?

514
00:22:58,820 --> 00:23:01,180
to get it to play nicer with an AI kind of infrastructure.

515
00:23:01,180 --> 00:23:03,200
Oh, yeah, I'm sure you need to, you know,

516
00:23:03,700 --> 00:23:07,100
adjust the software stack that drives it,

517
00:23:07,160 --> 00:23:09,460
but it's the same hardware.

518
00:23:10,100 --> 00:23:10,500
Nice.

519
00:23:11,380 --> 00:23:13,500
So I love this encapsulation of it all.

520
00:23:13,540 --> 00:23:15,020
The money chain is weirder than it looks.

521
00:23:15,740 --> 00:23:17,040
OpenAI calls your API.

522
00:23:17,740 --> 00:23:18,520
It runs on Azure.

523
00:23:19,180 --> 00:23:20,600
Azure leases from Equinix.

524
00:23:21,380 --> 00:23:22,480
Equinix is an REIT,

525
00:23:23,240 --> 00:23:24,760
so therefore your chat GPT query

526
00:23:25,420 --> 00:23:26,380
touches a real estate trust.

527
00:23:27,559 --> 00:23:28,600
Who saw that coming?

528
00:23:28,680 --> 00:23:29,320
Right.

529
00:23:30,940 --> 00:23:36,000
And anyway, listen, and we're not going to go way down the rabbit hole with the environmental impact. We kind of touched on that earlier,

530
00:23:36,240 --> 00:23:41,860
but just realize that large data centers can consume up to 5 million gallons per day.

531
00:23:42,080 --> 00:23:49,480
And that can be equivalent to the water use of a town of 10 to 50,000 people. And that water in this particular scenario,

532
00:23:49,800 --> 00:23:52,020
in most cases, it cools the gear,

533
00:23:52,240 --> 00:23:57,200
it evaporates up into the air, water cycle and lands somewhere else in the world.

534
00:23:57,260 --> 00:23:59,680
So essentially when you suck that water out of Northern Virginia,

535
00:23:59,940 --> 00:24:02,420
it may or may not come down again locally

536
00:24:03,140 --> 00:24:07,980
Yeah, and and we mentioned this when we talked about this once before but there's evaporative cooling

537
00:24:08,140 --> 00:24:14,740
There's closed-loop systems and the closed-loop systems are the ones that are actively the coolant is actively cooled

538
00:24:14,840 --> 00:24:20,440
It's not just water. Right and so they use dramatically less water, but a lot more electricity and

539
00:24:20,960 --> 00:24:22,720
Evaporative uses less electricity,

540
00:24:23,020 --> 00:24:25,019
but spends a lot more water

541
00:24:25,620 --> 00:24:31,440
Whichever one is greener depends on what the local environment can least afford to lose.

542
00:24:32,100 --> 00:24:32,220
So,

543
00:24:32,420 --> 00:24:37,020
you know, if the electricity in a region comes from clean sources,

544
00:24:37,340 --> 00:24:41,500
then that could be a motivation to use an actively cooled system.

545
00:24:41,960 --> 00:24:48,620
But I'm sure the problem is, you know, that's still going to be more expensive for the business that owns the data center.

546
00:24:48,720 --> 00:24:52,060
So they might not be as compelled to go that route.

547
00:24:52,680 --> 00:24:53,980
We really need fusion.

548
00:24:54,180 --> 00:24:54,760
We need it now.

549
00:24:55,200 --> 00:24:55,500
Please.

550
00:24:55,880 --> 00:24:56,460
Thank you.

551
00:24:56,720 --> 00:24:56,880
Totally.

552
00:24:57,840 --> 00:24:59,280
Which brings us to the nuclear option.

553
00:24:59,480 --> 00:25:01,040
We've mentioned this years ago,

554
00:25:01,220 --> 00:25:03,180
and it's not news necessarily,

555
00:25:03,510 --> 00:25:06,800
but these companies are maxing out local grids and taxing infrastructure

556
00:25:07,580 --> 00:25:13,160
to the point where they need to build their own or recommission nuclear power plants for their use specifically.

557
00:25:13,420 --> 00:25:15,500
Like Three Mile Island comes to mind for Microsoft.

558
00:25:16,080 --> 00:25:17,360
Right. Yeah, I wonder where that's at.

559
00:25:18,740 --> 00:25:20,120
Yeah, I don't know. I haven't heard much about that.

560
00:25:20,350 --> 00:25:22,100
I mean, I know where it's at. I just don't know.

561
00:25:22,140 --> 00:25:23,520
I know where it's at geographically.

562
00:25:24,840 --> 00:25:26,260
I was like, dude.

563
00:25:29,080 --> 00:25:30,680
You need me to draw a map for you.

564
00:25:31,059 --> 00:25:31,440
All right.

565
00:25:31,680 --> 00:25:33,020
It's Pennsylvania.

566
00:25:33,680 --> 00:25:33,740
Yeah.

567
00:25:33,960 --> 00:25:34,660
Right next door to you.

568
00:25:35,020 --> 00:25:36,820
Remember that being on the news back in the day when we were kids?

569
00:25:37,070 --> 00:25:37,500
Do you remember that?

570
00:25:37,780 --> 00:25:38,040
Oh, yeah.

571
00:25:38,420 --> 00:25:38,580
Right?

572
00:25:39,130 --> 00:25:39,860
1980 or something?

573
00:25:39,860 --> 00:25:41,100
What year was it?

574
00:25:41,440 --> 00:25:41,840
I think it was there.

575
00:25:42,700 --> 00:25:43,900
Have you driven past it?

576
00:25:44,020 --> 00:25:44,820
You can drive past it.

577
00:25:44,880 --> 00:25:45,440
Go over that bridge.

578
00:25:45,640 --> 00:25:49,120
I think it's over the Allegheny River or something, and you can see Three Mile Island down there.

579
00:25:50,260 --> 00:25:50,380
All right.

580
00:25:50,640 --> 00:25:51,620
Let's see what RAINA's got in the news.

581
00:25:52,460 --> 00:25:52,620
Thanks,

582
00:25:52,780 --> 00:25:52,900
boys.

583
00:25:53,620 --> 00:25:54,180
Erin Brockovich.

584
00:25:54,420 --> 00:26:03,820
The woman who took down a utility company for poisoning a town's water supply has launched a crowdsourced map to track what AI data centers are doing to communities across America.

585
00:26:04,600 --> 00:26:08,280
And the irony of her specific involvement here is almost too on the nose.

586
00:26:09,420 --> 00:26:15,720
Over 2,700 reports have already poured in, with Texas leading the misery parade at 612 complaints.

587
00:26:16,580 --> 00:26:17,360
And the top concern,

588
00:26:17,660 --> 00:26:19,480
you guessed it, is water.

589
00:26:20,360 --> 00:26:20,520
Apparently,

590
00:26:20,880 --> 00:26:23,840
civilization's AI cathedrals have a groundwater problem.

591
00:26:23,960 --> 00:26:26,920
and Julia Roberts is not available to fix it this time.

592
00:26:27,860 --> 00:26:28,960
People weigh this stuff,

593
00:26:29,140 --> 00:26:30,080
they weigh it differently

594
00:26:30,420 --> 00:26:32,780
and it depends on how close it is to their backyard,

595
00:26:33,820 --> 00:26:35,980
what local resources it threatens

596
00:26:36,400 --> 00:26:38,580
and then there's whether or not they believe

597
00:26:38,720 --> 00:26:41,300
that the AI arms race is real enough

598
00:26:41,460 --> 00:26:42,900
to make the sacrifices worth it.

599
00:26:43,600 --> 00:26:45,240
I'm not saying anybody should have it

600
00:26:45,480 --> 00:26:46,960
across the street from them.

601
00:26:47,320 --> 00:26:49,200
That just seems like a major impact

602
00:26:49,440 --> 00:26:50,580
on your quality of life.

603
00:26:51,380 --> 00:26:53,660
But I think they do need to be built.

604
00:26:54,940 --> 00:26:55,420
Yeah,

605
00:26:55,580 --> 00:26:56,460
they need to be built.

606
00:26:56,720 --> 00:26:59,940
And to your point, though, I do think the general public is,

607
00:27:00,460 --> 00:27:02,120
I wouldn't say becoming anti-AI,

608
00:27:02,340 --> 00:27:05,860
but I think they are now questioning the value of this

609
00:27:05,960 --> 00:27:07,020
at scale.

610
00:27:07,260 --> 00:27:10,260
You know, what is the reward and who is it for?

611
00:27:10,560 --> 00:27:12,540
It's really just benefiting a very

612
00:27:12,700 --> 00:27:14,080
small fraction of the population.

613
00:27:14,320 --> 00:27:18,219
I guess if you believe that whoever achieves super intelligence

614
00:27:18,220 --> 00:27:19,420
first wins,

615
00:27:19,640 --> 00:27:21,740
then that's a big factor.

616
00:27:22,490 --> 00:27:22,720
It is.

617
00:27:23,640 --> 00:27:24,800
And as you were discussing,

618
00:27:25,020 --> 00:27:28,960
mentioning all this, like I'm picturing, wouldn't it be somewhat

619
00:27:29,150 --> 00:27:33,220
obvious to park these things on abandoned stretches of coast near the ocean,

620
00:27:33,420 --> 00:27:34,040
build your

621
00:27:34,140 --> 00:27:35,260
own desalination plant,

622
00:27:35,500 --> 00:27:36,780
and then cool it that way?

623
00:27:36,890 --> 00:27:40,740
You know, I feel like there's, that will probably happen if it's not happening already.

624
00:27:41,720 --> 00:27:41,860
Yep.

625
00:27:42,020 --> 00:27:42,700
There goes another beach.

626
00:27:44,000 --> 00:27:44,440
Totally.

627
00:27:44,520 --> 00:27:45,200
Yeah,

628
00:27:46,160 --> 00:27:50,800
I know there might also be too much risk there for like, you know, flooding or, you know.

629
00:27:51,980 --> 00:27:53,900
Since we're on the topic of data centers,

630
00:27:54,160 --> 00:28:10,160
11 British MPs are pushing to amend the Cyber Security and Resilience Bill to give the technology secretary emergency last resort powers to shut down AI systems and data centers if they pose a catastrophic risk to national security,

631
00:28:10,500 --> 00:28:13,100
critical infrastructure or human life.

632
00:28:13,560 --> 00:28:17,100
Which is a very polite way of saying Parliament wants a hand on the plug.

633
00:28:18,100 --> 00:28:18,940
The catch,

634
00:28:19,120 --> 00:28:24,740
as experts immediately pointed out, is that modern AI is distributed across dozens of cloud regions,

635
00:28:25,020 --> 00:28:25,920
model layers,

636
00:28:26,180 --> 00:28:27,480
and third-party services.

637
00:28:28,440 --> 00:28:32,300
So shutting down one data center might just degrade the system rather than stop it.

638
00:28:32,840 --> 00:28:36,640
And any agentic AI that's already mid-task will keep executing regardless.

639
00:28:37,380 --> 00:28:41,340
It's the geopolitical equivalent of trying to turn off the ocean with a light switch,

640
00:28:42,070 --> 00:28:42,860
but points for effort,

641
00:28:43,040 --> 00:28:43,340
I guess.

642
00:28:45,140 --> 00:28:46,540
Are you a Lord of the Rings guy?

643
00:28:47,440 --> 00:28:47,520
Yeah.

644
00:28:48,480 --> 00:28:52,000
This immediately reminded me of, like, you see this meme all the time of Boromir.

645
00:28:52,780 --> 00:28:53,100
Yes.

646
00:28:53,510 --> 00:28:55,820
One does not simply unplug AI.

647
00:28:56,960 --> 00:28:57,880
Yeah, you're right.

648
00:28:58,570 --> 00:28:59,580
Two things come to mind here.

649
00:28:59,700 --> 00:29:01,880
First of all, I would park that switch in northern Virginia.

650
00:29:02,480 --> 00:29:06,920
I think I'd take a few data centers down there and maybe mitigate the risk somewhat.

651
00:29:07,200 --> 00:29:07,240
Yeah.

652
00:29:07,820 --> 00:29:10,120
And is it already too late?

653
00:29:11,040 --> 00:29:11,840
Is it already too late?

654
00:29:12,000 --> 00:29:12,060
Oh.

655
00:29:12,240 --> 00:29:14,540
This thing is already Pandora's box.

656
00:29:14,860 --> 00:29:15,560
Genie's out of the box.

657
00:29:15,560 --> 00:29:16,360
There's no way, right?

658
00:29:16,720 --> 00:29:16,880
Yep.

659
00:29:17,600 --> 00:29:18,680
We're all doomed.

660
00:29:19,620 --> 00:29:21,620
Starbucks deployed an AI inventory tool

661
00:29:21,650 --> 00:29:22,040
in September,

662
00:29:22,260 --> 00:29:25,740
2025 that promised 99% accuracy

663
00:29:26,290 --> 00:29:27,700
and the ability to count shelves

664
00:29:27,950 --> 00:29:29,360
eight times faster than a human.

665
00:29:29,800 --> 00:29:31,900
Then quietly killed it nine months later

666
00:29:32,100 --> 00:29:35,719
after the thing couldn't reliably tell oat milk from 2%

667
00:29:35,720 --> 00:29:37,980
and kept missing bottles sitting directly in front of it.

668
00:29:38,540 --> 00:29:41,220
The tool, built by Nomad Go using LiDAR sensors

669
00:29:41,340 --> 00:29:42,160
and computer vision,

670
00:29:42,360 --> 00:29:43,520
was a centerpiece

671
00:29:43,580 --> 00:29:46,200
of CEO Brian Nichols' "Back to Starbucks"

672
00:29:46,640 --> 00:29:47,520
turnaround strategy,

673
00:29:47,760 --> 00:29:49,000
which is now presumably

674
00:29:49,360 --> 00:29:50,700
back to humans with clipboards.

675
00:29:51,340 --> 00:29:53,600
In a masterpiece of corporate euphemism,

676
00:29:54,260 --> 00:29:56,480
Starbucks framed the shutdown as a move toward

677
00:29:57,220 --> 00:29:59,620
"consistency and execution at scale,"

678
00:30:00,060 --> 00:30:01,900
which is a very polished way of saying

679
00:30:01,940 --> 00:30:04,439
the AI couldn't count syrup bottles

680
00:30:04,440 --> 00:30:05,520
and the baristas were right.

681
00:30:06,560 --> 00:30:08,020
But that was nine months ago,

682
00:30:08,180 --> 00:30:08,640
this stuff.

683
00:30:08,840 --> 00:30:08,900
Yeah.

684
00:30:09,100 --> 00:30:12,360
I mean, I was actually surprised that it couldn't,

685
00:30:12,860 --> 00:30:15,500
I can't imagine Starbucks has an expansive inventory

686
00:30:15,960 --> 00:30:18,460
of wildly different things, not like a supermarket.

687
00:30:18,880 --> 00:30:20,900
I would think it would have been able

688
00:30:20,930 --> 00:30:24,340
to have been effectively trained on that data set

689
00:30:24,460 --> 00:30:25,560
and to do the job.

690
00:30:26,020 --> 00:30:29,080
But hey, that was September of 2025,

691
00:30:30,300 --> 00:30:30,520
man.

692
00:30:31,680 --> 00:30:32,240
Light years ago.

693
00:30:32,600 --> 00:30:34,320
I feel like you could vibe code this thing in your basement,

694
00:30:34,500 --> 00:30:35,760
like you, you know?

695
00:30:36,020 --> 00:30:37,260
That's next week's app.

696
00:30:38,040 --> 00:30:38,120
Yep.

697
00:30:38,740 --> 00:30:39,500
I'll sell it to Starbucks.

698
00:30:41,080 --> 00:30:43,540
AI is now designing the chips that power AI,

699
00:30:43,740 --> 00:30:45,840
which is either the most elegant feedback

700
00:30:46,060 --> 00:30:51,540
loop in tech history or a Ouroboros eating itself at $7.8 million per system.

701
00:30:52,380 --> 00:30:57,339
Google DeepMind's Alpha chip has already produced superhuman layouts for three generations

702
00:30:57,340 --> 00:30:59,140
of Google's own AI processors.

703
00:30:59,920 --> 00:31:01,700
And Berkeley researchers built a system

704
00:31:01,780 --> 00:31:04,300
that beat the state-of-the-art in processor cache design

705
00:31:04,720 --> 00:31:05,800
in just two days,

706
00:31:06,520 --> 00:31:08,700
which is the kind of sentence that should make chip engineers

707
00:31:08,940 --> 00:31:11,160
at least slightly nervous at their standing desks.

708
00:31:11,920 --> 00:31:14,420
The reassuring caveat from the researchers is that

709
00:31:14,540 --> 00:31:16,440
there is still a lot of human guidance,

710
00:31:17,220 --> 00:31:18,400
which, if you've been paying attention

711
00:31:18,600 --> 00:31:20,000
to how these stories tend to go,

712
00:31:20,560 --> 00:31:22,540
is exactly what people say in the paragraph

713
00:31:22,940 --> 00:31:24,500
before the humans stop being needed.

714
00:31:25,800 --> 00:31:26,760
Cyberdyne systems.

715
00:31:28,480 --> 00:31:29,780
Say it, man.

716
00:31:30,030 --> 00:31:31,000
You know you're thinking it.

717
00:31:31,120 --> 00:31:31,680
What was it?

718
00:31:33,360 --> 00:31:34,780
I can't even remember now.

719
00:31:35,800 --> 00:31:37,160
I'm visualizing that chip.

720
00:31:37,440 --> 00:31:39,520
Your Arnold has gotten much better, man.

721
00:31:39,820 --> 00:31:40,760
No, stop.

722
00:31:40,890 --> 00:31:41,260
You're sobbing.

723
00:31:41,260 --> 00:31:42,280
I'm sobbing.

724
00:31:42,520 --> 00:31:42,740
I'm sobbing.

725
00:31:43,920 --> 00:31:45,160
Got to the chopper.

726
00:31:46,780 --> 00:31:47,240
Different movie.

727
00:31:47,460 --> 00:31:47,520
Yeah.

728
00:31:48,000 --> 00:31:49,780
So, yeah, this was interesting.

729
00:31:50,000 --> 00:31:56,680
It made me feel like maybe that the path to self-improving AI might not be solely based in code,

730
00:31:56,900 --> 00:31:59,420
but also in hardware, too.

731
00:32:00,280 --> 00:32:03,500
And RAINA's so smart, I did have to ask her, what is an Ouroboros?

732
00:32:04,260 --> 00:32:04,900
Did you know what that is?

733
00:32:05,799 --> 00:32:10,080
Yeah, it's a snake eating its tail, but I had to Google it, too.

734
00:32:12,780 --> 00:32:13,820
And lastly,

735
00:32:13,980 --> 00:32:24,520
Anthropic and the Bill and Melinda Gates Foundation just announced a $200 million four-year partnership to deploy AI in the places the market was never going to reach on its own.

736
00:32:25,130 --> 00:32:30,360
Think vaccine research for polio and HPV, malaria forecasting for rural health workers,

737
00:32:30,600 --> 00:32:35,380
and math tutoring for kids in low-income countries who've never had a tutor in their lives.

738
00:32:36,030 --> 00:32:42,260
The target population is 4.6 billion people worldwide who currently lack access to basic health services.

739
00:32:42,560 --> 00:32:46,780
which is a number so large it's almost impossible to hold in your head.

740
00:32:47,480 --> 00:32:48,500
More than half the planet,

741
00:32:48,660 --> 00:32:52,380
effectively locked out of the healthcare system that the other half takes for granted.

742
00:32:52,960 --> 00:32:53,440
The kicker?

743
00:32:53,920 --> 00:32:56,700
Everything built along the way, the data sets,

744
00:32:57,000 --> 00:32:57,700
the benchmarks,

745
00:32:58,000 --> 00:32:58,460
the tools,

746
00:32:58,780 --> 00:33:00,740
gets released as public goods,

747
00:33:01,480 --> 00:33:03,380
which means this isn't just a charity flex,

748
00:33:03,640 --> 00:33:06,940
it's an attempt to build infrastructure that outlasts the press release.

749
00:33:07,920 --> 00:33:09,740
That's all the news for now. Back to you, gentlemen.

750
00:33:10,460 --> 00:33:14,700
But in exchange for that, we'd like to park this little data center in your backyard.

751
00:33:16,000 --> 00:33:17,200
There's an asterisk on that.

752
00:33:17,880 --> 00:33:18,640
I hope not.

753
00:33:18,720 --> 00:33:19,380
We need a fine print.

754
00:33:19,660 --> 00:33:19,880
I know.

755
00:33:20,300 --> 00:33:22,940
I hope there's a nice good news story to end on.

756
00:33:23,340 --> 00:33:23,620
Yes.

757
00:33:23,780 --> 00:33:25,520
Some altruism without strings attached.

758
00:33:25,800 --> 00:33:25,900
Yes.

759
00:33:26,060 --> 00:33:26,500
That'd be nice.

760
00:33:27,440 --> 00:33:28,220
Anything else, my friend?

761
00:33:28,640 --> 00:33:30,060
I think that's a wrap.

762
00:33:30,780 --> 00:33:31,260
All right, everybody.

763
00:33:31,460 --> 00:33:31,880
Thanks for listening.

764
00:33:32,040 --> 00:33:33,600
Subscribe on your favorite podcasting platform.

765
00:33:33,760 --> 00:33:34,560
Follow us on socials.

766
00:33:34,600 --> 00:33:35,040
Throw us a rating.

767
00:33:35,220 --> 00:33:35,920
We'll see you next week.

768
00:33:39,480 --> 00:33:40,700
This has been Up Against Reality.

769
00:33:41,290 --> 00:33:41,860
Thanks for listening.

770
00:33:42,640 --> 00:33:44,120
Subscribe to hear future episodes

771
00:33:44,440 --> 00:33:46,280
and be sure to follow us on social media

772
00:33:46,510 --> 00:33:47,400
for all things AI.

773
00:33:47,900 --> 00:33:50,400
Until next time, stay human, people.

774
00:33:50,480 --> 00:33:50,800
♪♪♪